Recent trends in marine microplastic modeling and machine learning tools: Potential for long-term microplastic monitoring

In 2020, it was estimated that 19–23 × 10⁶ tons of plastic waste enter the environment each year and it was found that plastic waste generation exceeds current mitigation efforts.¹ The global pandemic has led to an increased demand for plastics as many countries around the world implemented quarantining and other safety measures to mitigate the transmission of COVID-19. The use of equipment such as masks, gloves, COVID-19 antigen tests, and other medical supplies became widespread to increase public safety. Many of these necessities are made entirely from plastic or contain plastic components, leading to increases in plastic production. Daily lives were affected such that take-out containers, utensils, and other single-use plastic products were used. Additionally, many communities stalled or reversed policies that limited the use of single-use products, further increasing global plastic waste.² Already, masks and other COVID-19 associated plastic wastes have been found in the environment.^3,4 The transport and accumulation of plastics when they enter the environment can change further as they modified through fragmentation, photodegradation, biofouling, or coagulation, ultimately forming microplastics (plastic particles < 5 mm in size) and nanoplastics.^5,6 The increase in the global demand for plastics is a major concern for the future of plastic waste pollution and microplastics. Unforeseen changes in plastic consumption highlight the need for a better understanding of microplastic transport and fate.

Efficient microplastic monitoring is a major long-term goal necessary to understand the effects and developments of microplastic pollution. Microplastic research is labor-intensive and time-consuming,⁷ with current monitoring techniques requiring the collection, processing, and manual analysis of large quantities of samples, but these processes are not uniform or standardized.⁸ For effective monitoring to take place, it is also important to understand the patterns of microplastic transport to optimize the location and usage of developed monitoring tools. High-throughput and automated monitoring is also desirable for the efficient and large-scale analysis and understanding of microplastic accumulation patterns.^7–10 One potential avenue is to model microplastic behaviors to gain insight into the distribution, sources, sinks, and pathways.¹¹ Microplastic models are used for improved and realistic predictions of microplastic behaviors in aquatic environments. For example, it has been shown that models have the potential to optimize the locations in which plastic waste collection would be most effective across the globe.¹² 

While modeling can predict some behaviors of microplastics, it relies on empirical data about the microplastic behaviors and properties that are physically found and measured. Additionally, the complication of model equations is necessary to fully account for the complex dynamic forces governing inertial particle transport. Recently, machine learning and deep-learning models have been used to solve many problems involving large amounts of information autonomously, thus saving time and effort.^10,13–15 Image-based machine learning has been applied to material science studies that involve a large number of images to understand the structure and property relationships in materials.¹⁶ Yet, machine learning tools for microplastic analysis are still developing and there is room for improvement until their full potential is realized. Current machine learning and computer-vision tools focus on automating the counting and classifying of microplastics.^13–15 Classification is a processing step that can provide information about the size and shape of microplastics found in the environment. Depending on microplastic physical properties such as shape, density, and size, microplastics can exhibit different behaviors in the marine environment and can undergo different transport processes such as turbulent transport and settling velocity.^17,18 An innovative approach consisting of modeling and machine learning has potential to invoke major developments in microplastic monitoring and prediction of microplastic behaviors in the environment.

In this review, we outline the major features of available tools (i.e., models and machine learning), in the hope of inspiring studies that can bridge the gap between them to exploit their potential (Fig. 1). Numerical modeling can lead to a better understanding of microplastic distribution and most probable aggregation areas. A greater understanding of microplastic patterns can help researchers optimize monitoring and collection efforts. While this has involved satellite imagery for large plastic debris, some studies have also explored the use of numerical modeling to determine where to locate the smaller microplastics in the environment. Monitoring and collection efforts can provide valuable information about microplastic characteristics (size, shape, density, etc.), but accessing that information is challenging. Machine learning can aid in extracting such information from monitoring and microplastic collection more efficiently to provide larger amounts of data previously inaccessible. With more detailed information about microplastic characteristics, this information can be used to calibrate and validate the numerical models to further improve their predictive abilities. We will focus on recent studies regarding (1) transport models in which microplastic physical properties such as size, shape, and density are key parameters and (2) machine learning and computer-vision tools with the capacity to extract physical or chemical information from microplastics.

Understanding microplastics’ transport processes, including their complexities within the land to sea transition, is a grand challenge that must be overcome to understand and solve the microplastic pollution issue.¹⁹ The mechanisms underlying the transport of microplastics remain largely unknown because of the myriad of parameters that govern the physics of the problem. For instance, the distribution and transport of microplastics are influenced by the shape, size, density, and buoyancy as well as by ocean dynamics that interact with microplastics through waves, wind, and tides.¹⁸ Different categories of microplastics also have varying sinks, residence times, trajectories, and traveled distances.¹⁸ 

Numerical models are valuable in progressing the field of microplastic transport as they can be used to study the effects of different parameters (e.g., size, shape, density, and plastic type) systematically. Thus, numerical models help shed light on the fate and transport processes of microplastics, providing a holistic view of microplastics in the ocean environment (either locally or globally). When studying microplastic fate and transport, the models are classified based on the approach adopted to develop them, i.e., Eulerian and Lagrangian approaches.²⁰ While the former provides more information on the advection and diffusion phenomena from the hydrodynamics (velocity field) of the flow, the latter is more used to analyze the motions of the particles. According to Uzun et al., Eulerian and Lagrangian models can be further categorized based on their modeling method as hydrodynamic, statistical, mass-balanced, process-based, hydrodynamic and process-based, or hydrodynamic and statistical.²¹ Hydrodynamic models help analyze the influence of flow and physical properties of microplastics on the fate and transport of microplastics, allowing these models to simulate the advection and dispersion of microplastics in different geographies. Statistical models use Monte Carlo or Markov Chain analysis and are based on probabilistic hydrodynamic data or the probabilities of particle location.²¹ Mass balance and conceptual models for overall global transport can be useful however, are deficient in describing local catchment by neglecting factors such as rivers, flow distribution, geometry, time dependency, and turbulence.²² When considering the purpose of efficient microplastic monitoring, models that implement a hybrid approach, such as hydrodynamic and process-based and hydrodynamic and statistical model, are considered most informative and compatible with real-life findings.²¹ Hybrid models provide a more comprehensive understanding of spatiotemporal variations in microplastic concentration, sink regions, and horizontal and vertical transport processes.^11,18

Many numerical models are based on idealized physical laws and remain limited in that additional data are required to assess the occurrence and abundance of microplastics; broader particle shapes need to be included within the models.²¹ An added challenge to including physical properties of microplastics into models is that no universal shape descriptors of microplastics exist (see Fig. 2 for shape descriptors covered in this review). As such, the benefit of combining the findings of real-world microplastic studies and complex modeling requires further innovation. In Secs. II B–II D, we discuss some recent models that explore the effects of microplastic physical characteristics on microplastic transport processes.

An important property in microplastic transport is the velocity that particles reach during sedimentation, also known as the terminal settling velocity. Over time, most microplastics sink and accumulate in the sediment, which is considered the most suitable matrix to monitor and study the general composition of marine microplastics.²³ Previously, it was assumed that microplastics behave like sediments. However, in 2019, Waldschläger and Schüttrumpf determined that drag models that were previously designed for sediments are not appropriate for predicting the terminal settling velocity of microplastics.²⁴ About 500 physical experiments were conducted to examine the effects of size, particle shape (characterized by the Powers index and the Corey shape factor), and density on the settling velocity of microplastics. Adjustments to conventional sediment transport formulas were developed to describe the settling and rise velocities of microplastics based on each particle shape, and it was determined that there is a distinct difference between fibers and non-fiber particle shapes (e.g., spheres, pellets, and fragments).^24,25

A study conducted by van Melkebeke et al. aimed at determining which shape descriptors could be included in an empirical equation to describe and predict the sinking behavior of non-spherical particles in different fluids.²⁶ This study also investigated the effect of biofouling on polarity and how this contributes to the settling of microplastics and was the first to consider microplastic films as a morphology in a drag model. Physical plastics [high-density polyethylene (HDPE), low-density polyethylene (LDPE), polypropylene (PP), polyvinyl chloride (PVC), polyethylene terephthalate (PET), and polystyrene (PS)] of different shapes were used in sinking experiments. Several shape descriptors were evaluated, which included circularity, sphericity, Corey shape factor, power index, aspect ratio, flatness, and elongation. Despite many descriptors, it was found that no shape descriptor characterized and differentiated all particle shapes adequately. For example, sphericity was able to distinguish films from non-film particles while circularity was capable of distinguishing fiberous from non-fiberous particles. Biofouling was also determined as a contributor to the settling motion of low-density microplastics.

Biofouling occurs when organisms accumulate on the surface of objects and can cause microplastics to change their density, cause them to sink, and enhance their bioavailability.²⁷ Chubarenko et al. examined biofouling on slightly buoyant particles, using simplified physical and geometrical considerations, and assumed that the biofouling increased the total mass of the particle directly proportional to the surface area.²⁸ It was found that for the same mass, since many smaller particles have a larger surface area than one larger particle, larger plastics remained at the water surface longer.²⁸ Further exploring biofouling, Kooi et al. developed the first theoretical model simulating the effect of biofouling on microplastics and demonstrated that larger particles (10–1 mm) sink and resurface at a fast rate.²⁹ But smaller particles (<10 μm) sink slowly, which allows them to exist anywhere in the water column but then not resurface.²⁹ Comparing PE and PS microplastic particles in estuarine and coastal waters, Kaiser et al. offered the first empirical values to incorporate biofouling to microplastic sinking velocities.³⁰ The two different plastic types displayed a difference in sinking velocities in the two types of waters. For example, the sinking velocity of PS was greater in coastal water than in estuarine water, but for both waters, the sinking velocity decreased with lower temperatures and less light availability.³⁰ More recently, Jalón-Rojas has found that irregular biofilm distributions on microplastic fibers and sheets caused decreases in the settling velocity in sheets, despite an increase in the density, and modified the orientation of fibers resulting in increases in their settling velocity.³¹ 

A new shape parameter called the Aschenbrenner shape factor was used by Zhang and Choi to develop a shape-dependent drag model capable of distinguishing fibers, films, and fragments.³² Their focus was to determine a well-defined hydrodynamic shape factor, which includes contributions from the pressure drag coefficient and frictional drag coefficient to accurately predict the terminal settling velocity of fibers. The Aschenbrenner shape factor is considered a suitable hydrodynamics shape factor because it can theoretically calculate the contribution of pressure and frictional drag coefficients of an ellipsoid.³³ Interestingly, Zhang and Choi's model is based on the volume of microplastic fibers, which is rarely reported by microplastic studies. It is also important to note that the drag model is based on virgin microplastic data and how this drag model would compare to observations of real environmentally weathered microplastics is unclear.

Another model by Yu et al. derived a drag coefficient expression to predict the terminal settling velocity of microplastics based on the particle shape.³⁴ Such expression modifies the draw law for spherical particles (analytically derived from the laws of fluid mechanics) by including dimensionless particle diameter and two shape descriptors: sphericity and Corey shape factor. The corrected model reported the lowest errors (∼9% error) compared to the existing predictors that approximated particles as spheres. About 700 data points extracted from literature laboratory experimental data were used to develop the new drag coefficient expression, yet a limitation the authors mention is that more experimental data points are needed to expand the present model to laminar and turbulent regimes. Despite the low absolute error, the model lacks in representing the high fluctuations in the data possibly stemming from microplastic secondary movements such as rotation and tumbling that influence their sinking and increase the variation in the settling velocity and drag coefficient.

Secondary motions are oscillatory behaviors of particles that change settling orientation, which effects the drag coefficient and, thus, the transport of microplastic fibers. A drag model considering secondary motions along was developed by Choi et al. found that microplastic fibers initially accelerate before reaching a steady velocity and that microplastic fiber settling motions follow a sinusoidal pattern.³⁵ Additionally, a novel imaging processing technique was used to capture and characterize the settling of microplastic fibers. Microfibers ranging from 5 to 15 mm in length were cut from polyester ropes, placed into a water column, and captured with a camera creating videos at 1920 × 1080-pixel resolution and 50 frames per second. Both crosswise sphericity and the Aschenbrenner shape factor were considered because the former can account for settling orientation while the latter characterizes the unique morphology of fibers. While their drag model is useful, the size of the fibers is larger than environmental microplastics and only stiff and more cylindrical microfibers in non-quiescent water were considered.

Transport of microplastics along the coastal environments is the main process that influences the environmental fate and risks of microplastics.³⁶ In sea waves, the dispersion of plastics and microplastics is governed by not only particle shape but a series of interactions of many different processes, such as wind and waves.³⁷ In addition to vertical transportation in the water column, microplastics move horizontally by ocean dynamic forces. In oceanic transport, wind-driven mixing can cause plastic debris to distribute vertically within the upper water column.^38–40 Additionally, oceanic waves or currents can cause the selective transport of microplastics such that larger mesoplastics were trapped near-shore while microplastics were more dominant offshore.⁴¹ Modeling approaches can be useful in predicting where microplastics occur in the marine environment and are capable of evaluating locations to focus remediation efforts.^11,12,42 In this section, we highlight some efforts that incorporate microplastic physical properties to uncover the complexities of the horizontal movement of plastic and microplastics in the ocean.

The motion of particles in waves is a well-studied topic for many applications, including the transport and dispersal of microplastics. Surface waves impact the horizontal Stokes drift of particles and enhance the vertical sedimentation velocity of particles.^43–46 In 2018, DiBenedetto et al. studied the transport of anisotropic particles under waves and demonstrated that particle's shape (eccentricity) influences their preferred orientation and thus their settling velocity and transport.⁴⁷ This preferential orientation caused anisotropic particles to travel farther with less dispersal than randomly orientated particles, showing the importance of incorporating these findings in modeling microplastic transport.⁴⁷ Following up on these findings, DiBenedetto et al. also showed that this behavior is due to an interaction between the particle shape and the finite wave amplitude, the angular analog Stokes drift.⁴⁸ 

In 2019, Jalón-Rojas et al. developed a new framework to model the 3D transport of marine plastic debris called TrackMPD, a Lagrangian particle-tracking model, which shows that dynamical properties such as plastic density, shape, and size can affect plastic transport.⁴⁹ The study implemented the model for Jervis Bay (Australia) and focused on cylindrical and spherical particles and also included a wide range of physical processes such as advection, dispersion, windage, sinking, settling, beaching, and washing-off. The trajectories and fate of spherical particles with different densities and size were similar; however, denser particles settled quickly near the source, while the less dense particles were transported farther. Depending on whether spherical particles were low-density or intermediate, the particles were transported by different currents, suggesting that non-buoyant debris may travel longer distances than buoyant debris because of hydrodynamic mixing. Cylindrical particles on the other hand, regardless of density and size, traveled in similar trajectory patterns. The authors mention that estimation and prediction using the model requires accurate in situ measurements of the parameters used and additional microplastic samples are needed to assess particle properties. A more detailed review on TrackMPD and other well-known particle-tracking models (D-WAQ PART, Ichthyop, and CaMPSim-3D) is discussed by Bigdeli et al.²⁰ 

Besides understanding microplastic transport globally, modeling microplastic transport in a variety of local and regional cases can increase scientific understanding and initiate transformative changes. A case study conducted by Genc et al. applied a 3D coastal hydrodynamics, transport, and water quality model (HYDROTAM-3D) to study the movement of 3 mm sized polystyrene particles along coastal surface waters.⁵⁰ This study was mainly focused on the accumulation rate and microplastic concentration and did not consider plastic types other than polystyrene. Their model was applied in a case study for Fethiye Inner Bay in Turkey, part of the Mediterranean Sea, and found that microplastics preferentially accumulated in the south-west coastal waters, where the coastal circulation was the weakest, and east of the Murt River mouth (the major source of pollution discharge).

Another case study by Bondelind et al. was the first to investigate the influence of size and density of tire wear particles on the effect of their fate with hydrodynamic modeling, specifically in the Gota River in Sweden.²² It was found that particles with a size of 75 μm and a density of 1.9 g cm⁻¹ (larger and heavier microplastics) settled in the river most, only a third of microplastics with a size of 20 μm and a density of 1.7 g cm⁻¹ settled, and smaller microplastics with densities closer to 1 g cm⁻¹ did not settle and were able to be transported farther. However, other processes such as fragmentation and degradation assumed to be longer than the residence time of microplastics were not included in the simulations; therefore, the authors express the need to include more process-based modeling.

An important step in addressing global plastic pollution is estimating the amount of plastic in the ocean and developing techniques to understand accumulation and pathways. Because of the various processes influencing microplastic transport, it is challenging to study microplastic debris with data alone. Numerical modeling is a valuable tool for studying coastal processes through the simulation of physical, chemical, and biological features of coastal environments. Hydrodynamic conditions (i.e., waves, currents, wind-mixing) are temporal and can influence the transport of microplastics. Additionally, microplastics that undergo biological, chemical, and physical processes such as biofilm accumulation, aggregation, degradation, and sedimentation can alter particle densities and influence transport.²⁷ While hydrodynamic and process-based models are most comprehensive and compatible with real-life findings,²¹ it is a challenge to consider all processes and unique parameters that influence microplastic transport. Thus far, biofouling is a common process incorporated in transport models, and other processes such as fragmentation, which influences particle shape and requires longer times, could provide unique insights. The creation of a framework to accurately describe microplastic particle dynamics requires the inclusion of those processes that necessitate additional expansions to generalized model equations and increased computational complexity.^45,51

Integration of models and empirical observations can greatly improve the understanding of microplastics in the marine environment. Currently, there is a lack of standardization for plastic modeling and sampling methods that makes comparisons between models and sampling results challenging. Additionally, the lack of validation data can result in some discrepancies between observed and modeled values.²¹ Additional field data are useful in validating predictions for plastic accumulation hot spots, and larger datasets can aid in creating more robust models.²¹ This is especially significant as more studies are reporting the presence of smaller microplastics and nanoplastics.⁶ Microplastics come from a variety of sources with different sizes, shapes, and types. Shapes other than spherical particles such as fibers, fragments, and films are slowly being included, but these parameters vary much more in environmental conditions than what has been studied thus far. Models can also benefit from the inclusion of a wider variety of microplastic shapes.^21,46 Data from long-term monitoring are required to gain reliable information on microplastics’ loading, and the prediction of distribution in the marine environment¹⁸ and collection machine learning and computer-vision tools capable of high-throughput data collection (thus providing information about microplastic size, shape, and identity) may be helpful.

Providing large amounts of experimental data on microplastics to benefit the numerical model is challenging because there are few long-term studies regarding temporal variability in microplastic pollution and methods for microplastic monitoring are labor intensive. One of the most labor-intensive steps in microplastic analysis is the need for researchers to collect, identify, count, and classify microplastic particles. While visual identification is useful in microplastic classification (based on the shape and size), chemical verification is also crucial in understanding the scope of microplastic pollution and types of plastics that require further studying with numerical modeling. Chemical verification often involves vibrational spectroscopy techniques such as FTIR and Raman spectroscopy, but these techniques also require time-intensive analysis. This currently makes it impracticable for high-throughput and long-term monitoring because analysis requires the optimization of spectral collection parameters and expert analysis.

Microplastic research can benefit from reliable computer-vision based technologies that facilitate microplastic monitoring. Machine learning and computer vision are two fields that have become closely related to each other as machine learning provides new approaches to solving computer-vision problems and computer vision is a method to derive meaningful information from visual inputs. The use of computer-vision techniques and machine learning can lead to the automatic monitoring and collection of information of microplastics that are found in the environment. A strength of computer vision is the ability to objectively extract latent information from visual inputs and can characterize the shape factors of objects with regard to eccentricity, sphericity, elongation, etc., for various applications.^52,53 As more information about microplastics is collected, more accurate models can be developed, which ultimately results in more accurate microplastic transport predictions. Moreover, to fully understand the dynamics of plastic debris, there is a need to identify the physical characteristics and identity of plastic types¹⁸ and machine learning techniques can also be applied to microplastic spectroscopic analysis. In Secs. III B and III C, we discuss the emerging developments of machine learning and computer-vision tools for microplastic analysis.

In computer-vision techniques, segmentation is used to distinguish individual microplastics against a background to process physical feature information. One of the first applications of deep learning to microplastics conducted by Wegmayr et al. segmented microplastic fibers in digital images.⁵⁴ Lorenzo-Navarro et al. then developed a computer-based system to automatically count and classify microplastics 1–5 mm in size.⁵⁵ This system segmented microplastics from a RGB image of particles on a white background using a high-resolution flat scanner. For classification, different machine learning approaches were evaluated (k-nearest neighbors, C4.5, random forest, and support vector machine with radial based functions). The highest accuracy (91%) was achieved with a cascade classifier that included the particle shape, color, and texture features. In a follow-up study, Lorenzo-Navarro et al. also implemented a deep-learning approach with the aforementioned computer-vision system.¹³ Segmentation was performed with a U-Net network originally used for biomedical image segmentation, and classification was completed with a VGG16 network, which successfully classified microplastic shape with >95% accuracy. Compared to manual classification and the process developed in their previous study, their deep-learning method sped the process of classification by about 80% and 50%, respectively.

On a physically smaller scale, Shi et al. developed a deep-learning method to count and classify microplastics sized 50 μm to 1 mm from SEM micrographs.¹⁵ Image segmentation and shape classification was performed on a manually labeled SEM image data set of microplastic fibers, beads, and fragments. Semantic image segmentation was carried out using U-Net and MultiResUNet, achieving a Jaccard index of >0.75, where 0 signifies no overlap and 1 signifies perfect overlapping segmentation. Similarly, to Lorenzo-Navarro et al.,⁵⁵ Shi et al. implemented a VGG16 neural network for shape classification and achieved >98% accuracy. The benefit of using SEM images compared to microscopic images or optical cameras is that SEM images can provide more details about microplastic shapes, but at the disadvantage of longer sample preparation. The authors acknowledge that their model was trained on a small data set of images and should include larger more mixed datasets to identify and extract microplastic information for more complex environmental samples.

Machine learning on digital holographic images of microplastics also shows promise in identifying microplastics. Digital holography is an imaging technique in which a light wavefront of an object is recorded to create a hologram (consisting of two superimposed wavefronts creating an interference pattern). Machine learning approaches conducted by Bianco et al. used holographic images to determine holographic fractal fingerprints and identify microplastics and diatoms with fractal analysis.^56,57 From holograms, the authors implemented fractal (optimal mathematical descriptors) characterization using principal component analysis and trained a support vector machine to classify microplastics with >98% accuracy.⁵⁷ It was found that microplastics have smaller fractal dimensions, less lacunar, and larger fill ratio and vertex density compared to diatoms, which allowed the accurate discernment between the two populations using only 13 fractal features. Interestingly, the fractal analysis did not use size as an analytical parameter; therefore, the fragmentation of microplastics is not expected to influence classification. Since the chemical composition was not considered, their classification was also independent of size, shape, surface roughness, and chemical spectra. A potential advantage of this tool is that it could be used to achieve the automatic mapping of microplastics’ distribution in marine environments.

Besides counting and classification, another necessary step in microplastic analysis is chemical identification, which relies on spectroscopic characterization. Raman and FTIR analysis are two non-destructive vibrational spectroscopy techniques often used to identify the chemical material of microplastics.^58–60 However, spectral analysis is time-consuming and requires expert analysis to avoid misinterpretation.⁶⁰ To identify microplastics, the unknown spectrum must match with a known reference spectrum and technique is subject to some challenges in spectral interpretation. Many microplastics consist of a mixture of different polymers and additives that are not always included in reference libraries. Furthermore, Raman spectroscopy is particularly prone to fluorescence interference from additives and dyes that sometimes renders spectral identification impossible.^7,10,58,60 Additionally, environmental factors such as photodegradation and biofouling, can alter the spectra with the appearance or disappearance of spectral peaks.^5,60 While this is the case for both Raman and FTIR analysis, FTIR is more sensitive to oxidative groups from weathering and aging. To circumvent these challenges, some recent studies have developed machine learning methods to automatically interpret microplastic spectra.

A study conducted by Kedzierski et al. developed a machine learning algorithm using k-nearest neighbors for the high-throughput identification of FTIR spectra.⁶¹ A database with >900 microplastic FTIR spectra was created to automatically determine the spectra of microplastics. The algorithm was applied to 4000 spectra of unidentified microplastics and is highly dependent on the spectral database. While the method was efficient in identifying microplastics such as PE, PP, and polyamide (PA) less common microplastics were more difficult to determine. To find an alternative for spectral matching, Hufnagl et al. developed a supervised learning model using a random decision forest classifier to classify different types of polymers (PE, PP, poly(methyl methacrylate) (PMMA), polyacrylonitrile (PAN), and PS).⁶² However, there is an uncertainty whether the underlying particle is truly the identified polymer type.

The acquisition of high-quality spectral data is imperative for the chemical identification of microplastics and requires optimizing many parameters for analytical tasks; however, this becomes infeasible with high numbers of samples. Often for microplastic analysis, thousands of particles must be scanned to distinguish microplastic particles from other organic and inorganic particles. Brandt et al. applied an autoencoding neural network to reconstruct low-quality FTIR and Raman spectra.¹⁴ Autoencoders are a specific neural network architecture that first compresses input data (encoding stage) and then reconstructs the information into the original format with the network's decoder. The advantage of this technique is that even with low-quality spectra, the trained neural network can reproduce a clean spectrum for analysis. However, the network must be trained on a representative set of training spectra, so using the network to interpret μFTIR spectra is not possible after training the network on ATR-FTIR spectra.

In Raman imaging, a microplastic surface can be scanned over a pixel array (usually 10–10 000 pixels) to collect Raman spectral information as a matrix. Decoding large data matrices to an image to visualize microplastics and their chemical identity is time-consuming and can result in a loss of signal.⁶³ To overcome this, Fang et al. implemented principal component analysis to reduce the dimensionality of the data set and improve the interpretability of Raman spectroscopy for microplastic research.⁶⁴ The algorithm does not require Raman spectra directly for imaging; however, to verify the assignment of the imaged object as microplastic or other materials, the principal component analysis spectrum must be compared to the corresponding Raman spectrum.

While the aforementioned studies imaged and identified microplastics when already isolated and pre-processed, the ability to examine microplastics without pre-processing presents a benefit for researchers. Vidal et al. developed a fast microplastic identification technique that was capable of automatically identifying > 800 particles from sand based on near-infrared hyperspectral imaging and chemometrics.⁶⁵ Hyperspectral imaging creates a “chemical image” based on spectral information over different wavelengths from different spatial locations of a sample. The speed of near-IR hyperspectral imaging in combination with classification using soft identification of class analogy (SMICA) minimizes spectroscopic analysis as many particles can be quickly simultaneously analyzed and potentially identify microplastics as primary or secondary microplastics. This method has potential as a high-throughput screening technique although it would be more suitable for particles > 500 μm, so smaller particles still may benefit from FTIR and Raman instead.

Another challenge in microplastic analysis is the ability to detect microplastics in the water column because water strongly absorbs light in the IR region, which is generally used to characterize plastics.⁴² Huang et al. explored underwater hyperspectral imaging for in situ microplastic monitoring.⁶⁶ Identification explored various machine learning classification methods (with support vector machine exhibiting the best performance) to identify microplastics 0.5–5 mm in size, with various shapes, and colors. The significance of this study is that it first explores a simulated underwater environment using imaging equipment underwater to study microplastic abundance and has potential for the real-time analysis of microplastics in marine environments.

For large areas of marine litter monitoring, hyperspectral data from satellite imagery may allow for the better detection of larger floating plastic materials. While the ground-based monitoring of marine litter can provide precise information about the quality and quantity of plastic pollution, it can be time-consuming and expensive, and be difficult for larger-scale continuous monitoring. Apart from microplastics, larger patches of plastic debris can be detected through satellite spectral imagery as they have unique spectral signatures that can be distinguished from their surroundings. The Sentinel-2 satellites, launched by the European Space Agency, provide multi-spectral Earth observation data that can be used to identify large plastic litter for monitoring.⁶⁷ Using the Sentinal-2 satellite imagery, a few studies have developed models for the large-scale monitoring of marine plastic. In 2021, Jamali et al. evaluated random forest and support vector machine and the deep-learning method generative adversarial networks and random forest (GAN-RF) to Sentinel-2 satellite imagery.⁶⁸ Sentinel-2 satellite images were taken of plastic objects that were placed on the ocean surface. Compared to random forest and support vector machine classifiers, the GAN-RF classifier had a higher accuracy (96%) for detecting ocean plastics. Sannigrahi et al. also explored more functionality from the Sentinel-2 imagery and found that remote sensing bands and spectral indices are both important for developing a classification system to detect floating marine plastic.⁶⁹ Their random forest model was able to distinguish plastic with 91% accuracy.

More recently, in 2022, Taggio et al. presented a new machine learning approach to detect floating plastic from pansharpened hyperspectral data from PRISMA, a hyperspectral satellite launched in 2019 by the Italian Space Agency.⁷⁰ The study placed plastic targets on and offshore, from which PRISMA images were collected, and the offshore images were used as input data to train supervised (k-means) and unsupervised (light gradient boosting model, a decision tree method) machine learning algorithms to determine plastic spectral behaviors. PCA pansharpened images created higher spatial-resolution images of the high-spatial-resolution images taken by PRISMA to reach a spatial resolution of 5 m. Overall, it was found that a combination of both approaches was most capable of distinguishing floating plastic litter 0.6–5.1 m in size at about 96% accuracy.

Real-time analysis of microplastics is challenging, and effective microplastic monitoring demands a greater understanding of patterns of microplastic transport. Microplastic monitoring and prediction of their behaviors in the environment could benefit from innovation that involves numerical modeling and machine learning tools. Recent work in microplastic modeling has implemented the physical characteristics of microplastics, and machine learning tools are capable of extracting microplastic information. By incorporating physical parameters such as size, shape, density, and identity of microplastics, numerical models have predicted sinking velocities more accurately and estimated microplastic pathways in marine environments.

The issue remains that acquiring empirical data on microplastics is challenging and requires the application of tools that can handle the high-throughput necessary for microplastic analysis. Machine learning techniques have gained momentum in many fields and can solve the labor-intensive challenges that microplastic research presents. By providing more empirical data about environmental microplastics and potentially their behaviors, machine learning can feed much needed data to improve models. As models become more accurate, their ability to monitor microplastics improves, which aids in optimizing where to focus for more data sampling and collection. Data then can be used to train and broaden the capabilities of machine learning tools, thus forming a mutually beneficial connection between modeling and machine learning. Microplastic and plastic pollution has already been a global concern, but with the increase in production and use of plastics during the pandemic, the magnitude of pollution effects is uncertain. Undoubtably, the increase in plastic waste will result in an increase in microplastics in the future. Therefore, innovation and development of effective monitoring techniques is imperative to understand and manage the progression of microplastic pollution.

We acknowledge C. Hudson and A. Monti for thoughtful discussions and edits and the anonymous reviewers who provided critical feedback that greatly improved the manuscript. Financial support from the Okinawa Institute of Science and Technology Graduate University is gratefully acknowledged.

The authors have no conflicts to disclose.

Samantha Phan: Conceptualization (equal); Formal analysis (lead); Writing – original draft (lead). Christine Luscombe: Conceptualization (equal); Funding acquisition (lead); Supervision (lead); Writing – review & editing (lead).

The data that support the findings of this study are available from the corresponding author upon reasonable request.